Artificial Intelligence in medicine

(FL) is a groundbreaking technology in medicine that enables the training of AI models while maintaining patient data privacy. Unlike traditional methods that centralize data, FL keeps it at its source, such as hospitals. It involves distributing an AI model to various nodes, where each node trains it locally without sharing raw data.

Only updated model parameters are sent to a central server, which aggregates them to improve the model, ensuring compliance with strict data protection regulations.

Applications in Medicine

1. Medical Imaging Diagnosis

FL is particularly useful in radiology and pathology, where datasets are often enormous and scattered across different institutions.

• Brain Tumor Detection: A FL model can be trained on MRIs from various hospitals. Each hospital trains the model with its own images of tumors and healthy patients. By combining the knowledge from each center, the final model can detect tumors with greater accuracy and generalize better to different types of scanners or patient populations.
• Pneumonia Detection from X-rays: Similarly, a model can learn to identify signs of pneumonia from chest X-rays from multiple clinics, improving the model's robustness by being trained on diverse representations of the disease.

2. Precision Medicine and Pharmacogenomics

FL can be used to develop models that predict a patient's response to a medication or the risk of a disease based on their genetic profile.

• Predicting Optimal Drug Dosage: Various hospitals can collaborate to train a model that predicts the correct dose of a drug for a patient, using genomic and clinical data, without the need to share confidential patient information.
• Biomarker Identification: FL allows for the collaborative analysis of large genomic and clinical datasets to identify biomarkers associated with certain diseases, such as cancer.

3. Mobile Health and Wearable Devices

FL can be used to train models that analyze data from wearable devices (smartwatches, activity monitors) to track patients' health.

• Arrhythmia Prediction: A model can be trained using electrocardiogram (ECG) data collected by mobile devices from thousands of users. FL allows the data to remain on the user's device while the model learns to detect patterns of arrhythmias, such as atrial fibrillation.

Comparison with Other Technologies

FL represents a paradigm shift that allows AI to reach its full potential in healthcere, overcoming privacy barriers and enabling large-scale collaboration. It is a key tool for building more robust, equitable, and useful AI models for healthcare.

We have developed a federated learning platform designed for privacy-preserving AI

We have several partnerships and use cases in the healthcare sector, which they are highlighted on our website and in various publications.

Here are some examples of how our federated learning platform is being applied in healthcare:

1. Advancing the Diagnosis of Rare Diseases

This is a key example in healthcare. We worked with the National Institutes of Health (NIH) and University College London (UCL). Their goal is to improve the diagnosis of Collagen VI-related dystrophies (COL6-RD), which is a rare genetic disease.

• The Challenge: Diagnosing rare diseases is difficult because patient data is scarce and scattered across different hospitals and research centers globally. Centralizing this sensitive medical information would violate privacy regulations like HIPAA and GDPR.
• Our Solution: The platform enables the NIH and UCL to collaboratively train a machine learning model on collagen VI immunofluorescence microscopy images. The data never leaves the institutions. Each organization trains the model with its own images, and only the model updates are shared and aggregated.
• The Outcome: By training on a larger, more diverse dataset, the model's diagnostic accuracy is significantly improved, without a single patient record being transferred. This marks the first global application of federated learning for COL6-RD diagnosis.

2. Collaborative AI-Assisted Diagnosis

Our AI platform is used to enable collaboration between hospital networks to train shared diagnostic models for more common conditions.

• The Challenge: Hospitals often have their own proprietary datasets for diseases like lung cancer, diabetic retinopathy, or Alzheimer's. While valuable, these isolated datasets may not be large or diverse enough to create a highly robust and generalizable AI model.
• Our Solution: Multiple hospitals can use the platform to jointly train a disease detection model. Each hospital trains the model locally using its own x-rays, retina images, or patient records. The platform securely synchronizes the model parameters.
• The Outcome: The resulting model is more accurate and effective because it has learned from a broader range of data, including different populations, equipment, and disease presentations, all while keeping patient data private and secure within each hospital's firewall.

3. Personalized Medicine and Treatment Plans

The platform facilitates the development of patient-specific predictive models.

• The Challenge: Precision medicine requires analyzing vast amounts of clinical, genomic, and pharmacological data to predict a patient's response to a specific treatment. This data is highly sensitive and exists in different data silos.
• Our Solution: Hospitals and research centers can collaborate to train predictive models for immunotherapy response in cancer patients, for example. The model learns from data from diverse populations in different institutions without the sensitive genetic or health data ever being moved or shared.
• The Outcome: This allows for the creation of therapies tailored to individual patients, while maintaining full data control and regulatory compliance.

4. Accelerating Clinical Trials

Our technology is also used to accelerate medical research and clinical studies.

• The Challenge: Multicenter clinical trials often involve transferring patient data across international borders, which is a slow and complex process due to varying privacy regulations.
• Our Solution: A decentralized clinical trial on the effectiveness of a new drug can train a model on distributed data from multiple centers without transferring patient records across borders.
• The Outcome: This streamlines the research process, reduces time to market for life-saving therapies, and ensures compliance with local privacy regulations.

In all of these cases, Our Federadet AI platform emphasizes the key principles of federated learning: data sovereignty, privacy by design, and the ability to build a collective intelligence from decentralized data sources.

Also incorporates features like differential privacy to add an extra layer of security, mathematically ensuring that individual data points cannot be reconstructed from the shared model updates.

When it comes to the intersection of federated learning and genomics, several research papers and platforms are exploring this promising field. The challenge in genomics is particularly acute, as genetic data is highly sensitive and often siloed in different institutions, making large-scale, collaborative research difficult.

Here are some key research papers and platforms related to federated learning in genomics:

Efficacy of federated learning on genomic data: a study on the UK Biobank and the 1000 Genomes Project

This paper, published in Frontiers in Big Data, provides foundational work on the applicability of federated learning to individual-level genomic data.

• Key Contribution: The study directly investigates whether a federated model can achieve performance comparable to a centralized model when predicting phenotypes from genotypes and ancestry from genomic data. They use two well-known and massive datasets, the UK Biobank and the 1000 Genomes Project, to conduct their experiments.
• Findings: The researchers demonstrate that federated models trained on data split into independent "nodes" (simulating different data silos) achieve performance very close to a centralized model. This is a crucial finding, as it proves that the privacy-preserving nature of FL does not necessarily come at the cost of model accuracy in genomics. The paper also explores the effects of communication frequency between clients and the central server on model accuracy.

Breaking Down Barriers in Global Genomic Research

This is a review article published in the journal Genes, which provides a comprehensive overview of the potential and challenges of FL in genomics.

• Key Contribution: This paper details the methodology of FL in the context of genomic research, explaining how institution-specific genomic data (like FASTQ, BAM, or VCF files) can be used to train local models. It discusses the process of secure aggregation and iterative model improvement.
• Discussion: The authors examine key challenges, such as integrating heterogeneous data from different sequencing platforms and the cybersecurity risks associated with FL. They also discuss how FL can help with compliance with regulations like GDPR. The paper highlights successful implementations and discusses future directions, including the integration of FL with other AI techniques to advance precision medicine.

Democratizing clinical-genomic data: How FL platforms can promote benefits sharing in genomics

This perspective article, published in Molecular Systems Biology, explores the role of federated data platforms in making genomic data more accessible for research while promoting "benefits sharing" with participants.

• Key Contribution: The paper argues that federated data platforms are a means to achieve data accessibility and security simultaneously. It describes how these platforms can be used to virtually link "trusted research environments" (TREs) where patient data is stored, without physically moving the data.
• Discussion: The authors discuss the governance and ethical considerations of these platforms, emphasizing the need for strict regulations and accreditation schemes to build public trust. They highlight that a key benefit of federation is democratizing access to global data assets and improving the representation of underrepresented groups in genomic research, as sequencing efforts have not always been diverse.

Federated Learning Enables Big Data for Rare Cancer Boundary Detectio (from the FeTS project)

While not exclusively focused on genomics, the Federated Tumor Segmentation (FeTS) project is a prime example of a large-scale, multi-institutional federated learning platform in oncology, which heavily involves genomic and clinical data.

• Key Contribution: Published in Nature Communications, this paper details a collaborative effort involving researchers from 6 continents and 71 sites to train a model for glioblastoma (a rare brain tumor) segmentation.
• Findings: By using an FL approach, the researchers were able to leverage a large and diverse dataset of over 6,300 patients without needing to move, integrate, or harmonize the data. This significantly improved the model's accuracy and generalizability compared to models trained on smaller, single-institution datasets, demonstrating the power of FL for rare diseases where data is scarce and distributed.

Our AI platform allows for secure

Collaborative AI model training on genomics data without requiring the data to be centralized or shared. This capability is crucial because genomic information is highly sensitive and often siloed due to strict privacy regulations and institutional firewalls.

Here are the key abilities of aour AI platform in genomics:

Privacy-Preserving Research Collaboration

The platform's primary advantage is its ability to facilitate joint research between multiple institutions—such as hospitals, universities, and research centers—without compromising patient privacy.

Instead of pooling raw genomic data, the platform sends the AI model to each institution's data silo. The model is trained locally, and only the encrypted model updates are sent back to a central server. This process is repeated until a global, highly accurate model is created.

• Practical Example: The company has a notable partnership with the National Institutes of Health (NIH) and University College London (UCL). They used the platform to train a model for diagnosing Collagen VI-related dystrophies (COL6-RD), a rare genetic disease. By training on microscopy images from both institutions, the model improved its diagnostic accuracy without any patient data ever leaving its source.

Overcoming Data Scarcity for Rare Diseases

Genomic data for rare diseases is inherently scarce and fragmented across different locations globally. The platform helps to overcome this challenge by enabling researchers to collectively build a larger, more diverse training dataset virtually.

• Benefit: This allows for the creation of more robust and generalizable models than would be possible for a single institution. For diseases where a few dozen cases exist worldwide, federated learning can be a game-changer by unifying the "knowledge" of these scattered cases into a single, powerful AI tool.

Regulatory Compliance

Genomic data is subject to some of the strictest data privacy laws, including HIPAA in the U.S. and GDPR in Europe. Our Federadet solution is designed with these regulations in mind.

• How it works: Because the raw data never leaves the institutional firewall, the platform significantly reduces the risk of data breaches and simplifies the legal and administrative complexities of cross-border data sharing, ensuring full regulatory compliance.

Enabling Personalized Medicine

The platform can be used to develop predictive models for personalized medicine. These models can analyze genomic, clinical, and lifestyle data to predict a patient's risk of disease or their likely response to a specific drug.

• Application: Researchers can collaboratively train models to identify genomic biomarkers associated with a particular disease or a treatment's effectiveness, without each institution having to share sensitive patient information. This helps accelerate the development of personalized therapies and treatment plans.

Unlock the power of data & AI 

Don't wait more, book a demo with us, one of our experts will provide a platform demonstration tailored to your needs, including recommendations to confidently implement our solution to drive value for your organization.

• Define your custom use case
• Work towards a better use of your data
• Train your model while preserving privacy